Research On Object Detection Algorithm Based On Weakly Supervised Learning

Object detection is mainly divided into two forms: strongly supervised learning and weakly supervised learning.There are some problems in strongly supervised learning,such as the high cost and difficulty of object labeling.At present,there is no good solution.Weakly supervised learning can greatly alleviate the problems in strongly supervised learning.However,there are some problems in weakly supervised learning,such as incomplete,inaccurate and inaccurate label labeling,As a result,the average accuracy of correlation algorithm in the application of object detection is low,and it can not complete the task of object detection better.In order to solve the problems existing in weakly supervised learning,this paper proposes a weakly supervised object detection model which combines domain transfer technology and Continuation Multiple Instance Learning algorithm.Continuation Multiple Instance Learning algorithm is an improved weakly supervised learning algorithm based on multi instance learning algorithm.First of all,Cycle GAN domain transfer technology is used to transform the source domain image data set into three different forms of data set images: Comic,Watercolor and Clipart,which are similar to the object domain images.Then,the object region in the detected image is extracted by convolution neural network,and the instance label is estimated by the two modules of instance selection and instance classification.In the pro cess of instance label estimation,in the object instance contained in the sample data image,multiple object bounding boxes usually appear for a single object,This leads to the inaccuracy of the boundary box of object detection.To solve this problem,th is paper adds a Non-Maximum Suppression algorithm to eliminate the redundant object bounding box on the object instance,and takes the object detection bounding box with high confidence score as the pseudo object annotation of the object instance.Finally,using the obtained domain transfer image and pseudo object annotation image,the pre trained object detector is fine tuned in turn to obtain the final object detection model.Compared with the original weakly supervised object detection model,the improved object detection model proposed in this paper can improve the accuracy of pseudo object labeling to a certain extent,and significantly improve the average accuracy of the object detection model based on weakly supervised learning.